Hand tracking v13 — Point cloud
Segmentation keeps only the person, rebuilt as a hovering cloud of coloured dots whose depth comes from a swappable rule: pixel value, wave, noise or motion.
Only the person survives the picture: where live segmentation finds a body, a grid of coloured dots hangs in space, and everywhere else is void. The cloud's surface bulges and recedes as its depth rule works, while a slow camera wobble adds just enough parallax to make the relief legible.
The source frame is downsampled to one pixel per grid cell — a free colour average — while the segmentation mask decides which cells exist at all. A swappable mode then turns each cell into a depth: the pixel's perceived value, an animated wave, a Perlin field, or frame-to-frame motion that pushes only where the subject moves. Every point eases toward its target so the surface never crackles, and the dots render into a 3D scene sized by their depth, coloured from the footage or a spectral palette.
